THE METHODS OF EXTRACTING WATER INFORMATION FROM SPOT IMAGE

CHINESE GEOCRAPHICAL SCIENCEVolume 12. Nunber 1. pp. 68 - 72. 2002Srience Press. Beiing. ChinaTHE METHODS OF EXTRACTING WATERINFORMATION FROM SPOT IMAGEDU Jin-kang, FE.NG Xue-zhi，WANC Zhou-long，HUANG Yong-sheng, RAMADAN EInazir( Department of Uirban and Resoure Science, Nanjing Unirersity. Nanjing 210093, P. R China)ABSTRACT: Some lechniques and methods for deriving water information from SPOT - 4(XI) image were investigaltedand discused in this paper. An algorithm of deision-lree (DT) casication which includes several casifers based onthe spetral responding characteristics of water bodies and other a山jects。was develuped and put forward to delineate WH-ter bodies. Another algorithm of decision-lree casfication based on both spectral etion of DEM and slope (DTDS) was also designed for water bodies extraction. In addition, supervised lasificationmethod of naximum-likelyhoud cassification (MLC), und unsuperised method of interactive sll-organizing dada analy-ais techniqpue (ISODATA ) were used lo extract walerbodies for cormnparison purpose. An indexs was designed and used toa5s15s the aceuracy of differenl methods adopted in the rcsearch. Results have shown that water extraction accuracy wasvuriable with respect to the various lechniques applicd. It was low using ISODATA, very high using DT algorithm andmueh higher using loth DTI)S and MLC.KEY WORDS: water lordy; decision tree algorithm; acuracy assesementCLC number: TP751. I .Document eode: A .Article ID: 1002-0063(2002)01-0068-05Extraction of water information from digital salelliteIB VHRSR dala, they indicated that waler bodies couldimages has been studied broadly in recent twenty years. be identified if the ratio of CH2 and CHI be used.The methods of identifying waterbodies such as thresh-ZHOU Cheng-hu et al. (1996) and Du Yun-yan et al.olding, segrent, I andsat chromaticity coordinates,(1998) developed a descriptive model for automaticallyproportion estimation. and descriptive algorihm based extracting and recugmizing water bodies based on theon knowledge of water spectrum feature, have been putknowledge of water spectrum feature using NOAA/forward and applied to a variety of satelite data for AVHRR data. YANC Chun-jian et al. (1998) de-managenent of water resources and monitoring of signed a algorithm to extract waler bodies from Landsatfloouls. SHIH (1985) used Landsat MSS data to delin~ TM. Based on their analyzed results that the sum ofeate the water surface area from the surounding land.TM2 and TM3 were larger than that of TM4 and TM5 forIn his research he indicated that both techniques ofwater bodies, they used the algorithm to distinguishdensity slicing from band7(near infrared) and ELAS waler bodies from shadows eelctively in mountainousclassification with combination of bands 5 and 7 could areas. Other methods were used to extract water infor-successfully assess the waler-surface area. The devia~ mation frum different salellite data By BARTON IJ eltion of the surface area assessment belween two tech-al. (1989),LIU Jian-bo et al. (1996)，XIAOniques was within 3%. LU Jia-ju(1992) used tech- Qian-guang et al. (1987). Due to the different spatialniques of thresholding, Landsat chromaticity coordi-resolutions of satellite data and geographical character-nales and" proportion estimation" to extract water bod-isties of study area, a method or lechnique can only bees based on Landsat CCT data, and found that the adopled based on the deep analysis of salite data, theproportion estimation approach can distinguish smallerphysical-reosranhical fpatures of the study area and thewater bodies efectively. SHENG Yong-wei et al. spectral中国煤化工ther objects. In(1994) tried tw discriminate waler bodies using FY一this papHCNMHGingwalerinfrRerceived dale: 2001-11-15Biography: DU Jin-kang(1964-)，male. a native of Muping County, Shandong Province, associate professor. His research interestsinclude hydrological modeling, application of GIS und Remote Sensing l0 bhydrology.The Methods of Extracting Wauter lnformation From Spot Images9mation from SPOT-4 data were investigated. An al- bodies effectively frorm image in mounlainous area, bothgorithm of decision-tree (DT) lasification with several information from infrared and visible bands could beclassifers based on spectral responding values was de- used lo identify the difference between waler bodies andsigned to derive waler infurmation. Another algorithm of shadows.decision-tree classification algorithm based on bothIn SPOT image of the study area. five typical landspectral responding values and auxiliary information of cover classes were determined and training samplingDEM and slope (DTDS) was developed. Supervised were taken to calculate means and standard deviations ofclassfication method of maximum- likelyood lasifica- their spectral responding values (Sld. Dev.) (Tabletion (MLC) and unsupervised classification method of 1)， which are re-sampled DN values. The coincidentinteractive self-organizing dada analysis technique spectral plot (mean plus and minus two standard devi-(ISODATA) were also used to extract water informa- ations for the five types in each band) was drawn intion in the same area. The results and accuracy of the Fig. I. In this figure it was clear that the ranges ofmethods were compared and evaluated.spectral values of water and other objects were over-lapped in B1 and B2. In B3 the range of waler spectralI THE SPECTRAL RESPONDING CHARACTERIS-values also overlapped with that of shadows, but theyTICS OF WATER BODIES .were lower than that of other three classes and shadowshave litle overlapped area with plants. There was aThe Jiangning County of Jiangsu Province was se- clear distinction between the group of waler, shadowlected as study area and the SPOT-4 image( XI，and the remaining three classes in SWIR, but there is1999.1 ) with four bands of B1(0. 59-0.59μm)，B2 still a big overlapping area between water and shadow.(0. 61 - 0. 68μm)，B3(0. 78 -0.89μm)，and SWIR Based on the analyzed spectral characteristics of five(1. 58 - 1.75μm) was used for the study.land cover elasses in SPOT (XI) image, the ollowingThe original SPOT-4 image was rectifed with different approaches were used for extracting water in-known CCPs obtained by GPS and from relief map of formation and the resuls were evaluated.1: 50000. The nearest neighbor interpolation algorithm25was used to re-sample the Digital Number( DN) value ofeach pixel.200The image mainly recorded the information of re-n-rngatiedd fneldlection and radiation of the objects. For the differentresidential aicastructures. components， physical and chemical char-acteristics of objects, the retflection and radiation ar10varied in the electromagnetie wave bands. In naturalcondition, even when the water is very shallow，water50. pinmibodies absorb nearly all incident energy in both theshadennear- -infrared and middle-infrared wavelengths, andSWRthere is very lttle energy available to be relcted. Thatbandsleads to water feature having a significant and distinctlylower reflectance than either vegetation or soil through-Fig. 1 The coincident spectral plot of fivetypical spectral elassesout the reflective infrared portion of the spectra. In in-frared image, and for the same reason, water appearsdark while soil and vegelation appear bright. Thus it's2 THE METHODS OF EXTRACTING WATER IN-easy to derive water form other objects using threshold-FORMATION FROM SPOT IMAGEing in infrared band. Unfortunately， in mountainousarea，the objects under the shadow of mountain also2. 1 The Decision Tree Method Based on Spectral Val-reflect lttle energy and appear dark in image, it's dif- ues(DT)ficul 10 exlract water bodies by the use of thresholdingin infrared band. In visible light band, the reflectedI 中国煤化工aer were le lhaninformation of water in image mostly come from mattersthos，the thresholdingin water surface，inside waler, and at water bottom,techMHCN M H G water information.which could be used to recognize the deepness, quality The histogram ol SWIK was bult to determine theand bedload content of water. In order to extract water threshold. It is clear from the histogram that there were70Du Jin-kang, FENG Xue-zhi, WANG Zhoulong el al.Table I The statistic index of samplesB2_BSWIRTypemeanId.maptd.Std.eyDeWater108243440Plar986241349316Residential area105810112Shadow1014836Non- irigated field111711067two peaks belween water and other objects, and smoothwater.ixells were obvious and few shadow pixels weredip between the peaks. The threshold value of 85 was found in water image by flicking the waler image andselected ，which was litle close to the edge of the peakcolor composition image of B2， B3， SWIR. Theof other objects so that the water would not be losl. The steps of extracting water bodies could be laken as de-extracted water image using thresholding was overlaid oncision tree algorithm (Fig. 4), at each step a classi-the color image of B2，B3, SWIR. By flicking the two fier was designed, and more pure classes were ob-images, we found that nearly all water pixels were se-tained.lected, Some of them were not belong to water, but tothe shadows of mountains(Fig. 2).Mounuainous shadowsFig. 3 Water image extracted using DTFig. 2 Water imege extracted using thresholding in SWIRImput Pixe!In order to discard shadows from the image of ex-tracted water, sampling o{ water and shadows weref SWIR<85>_ NO Non.wata Pixelaken from the water image. It could be seen fromYESpectral responding curves of the two types of samplesthat most spectral values of water pixels were greater<自100 NO≤E1283. B2>60. B3<55 SWIRS0Pthan that of shadow pixels in B1. The threshold tech-YES .nique could be adopted to discard shadows, nearly allWianer PixelShasdo4 Pixe]shadow pixels were discarded if threshold level of 100was used, but some of water pixels were lost.Fig.4 The decision tree diagam of extacting waterTo withdraw the lost water pixels， water anshadow samples on discarded shadow image were col-lected to find if the difference of spectral responding 2. 2 The Decision Tree Method Based on Both Spectralexist between water and shadows. It was clear from theValue and Auxiliary Information of DEM and Slopeanalysis that the spectral values of most water pixels(DTDS)were: 1) great 83 in B1; 2) grealt 60 in B2; 3) less 55in B3; 4) less 50 in SWIR. The lost waler pixels couldAuxiliary information such as DEM and slope canbe taken out from shadows image if they met the fol-be中国煤化工n land cover/ uselowing set of conditions:classlese auziliary infor-B1> 83 and B2> 60 and B3 <55 and SWIR <50matiqMYHC N M H G using decision treeThe water image obtained through the above stepsalgorithm.vas shown in Fig.3. It could be seen that nearly allThe Methods of Extracting Water Informnation Frorm Spot lmage7Thresholding technique could be used lo withdrawnique (ISODATA) was adopted lo elassify the informa-water pixels from the image by using the threshold value tion of SPOT XI into 15 classes. The waler, as one ofof 85 in band SWIR. The water image have had somethe classes, was withdrawn.shadow pixels needed to be discarded. The DEM ob-tained from topographic map and slope derived from3 ACCURACY ANALYSISDEM are added to the water image as other auxiliarybands. Sampling was taken to analyze the differences o{Usually，accuracy assessment is done through aDEM and slope between water and shadows pixels. Itcomparison of test pixels. Sample pixels are selectedwas found that most DEM values of water pixels were within a square window using one of the random, strat-less than 110 and the values of slopes less than 6,ified random and equalized random stralegy. Kappatherefore the elassifier of DEM <110 and slope <6cofficient is derived from confusion matrix produced bycould be used to separate the water image into water and comparing the classification results for the test samplesshadow ones. The procedure of the method was shown invith the reference data. For accuracy assessment ofFig. 5.waler extraction methods, only two types of pixels(water and non-water) were contained on the thematicwater image, when non-water pixels occupy a large partsInpul Pixelof the image, the estimated Kappa cefficients for allSWIR<85> Non wacrPixel0methods would retain high values with lttle differences,leading to the failing to renlect the accuracy of eachyesmethod. A new assessment approach was developed toattempt to evaluate each method effectively. The pro-DEMCII0& slopescedure of the assessment was to compare the referenceNCwater image with those extracled using other methods.YESthroughout all pixels on the image. Generally. waterWalter PlixelShadowg Piexlsurface area varies with lime, it is difficult to obtainground truth data about water area. The method wasFig, 5 The decision tree diagram of extracting water withsupposed to be with bigh accuracy, if the water bodiesBuxilary information of DEM and slopeextracted using the method have a good visual agreementwith original image. The DT method achieved best re-2. 3 Supervised Classification Method of Maximumsults by comparison, therefore, the results obtained byLikelyhood lassification( MLC)DT method can be used as the reference water image inaccuracy asssment. The aceuracy of each methodIt was easy to find visually the differences betweencould be assessed by calculating the user-accuracy co-water and shadows in stretched waler image derived eficient K and computation accuracy cofficient C.sing threshoding technique in band SWIR, in whichwhich were defined as follows:some shadow pixels existed. Therefore, the method ofK = WATER/ (WATER + WATER+)maximum likelyhood classification(MLC) could be usedC= 100-(WATER- + WATER+)/W ATER(REF)to classify the water image. The block-training strategywhere: WATER - number of pixels which are labeledin which a number of contiguous pixels were used as thewith waler both on reference map and compared map;training samples was employed for training. Afterw. ATER-一number of pixels which are labeled withtraining， statistics (signitures) of water and shadowwater in reference map and non-water in compared map:were obtained. The water and shadow were classifiedWATER+- number of pixels which are labeled withusing the MLC method, and the water bodies were ex-non-water in reference map and water in compared map;lracted.WATER(REF)一rumber of pixels which are labeledwith water in the image obtained using DT. Table 2 was2. 4 Unsupervised Classification Methodthe resuls obtained by comparing all pixels on referenceTesults showed that theThe easiest way to extract water bodies without any hig |中国煤化工odies from SPOT(XI)spectral analysis is the use of unsupervised classification:ouCN M H Ghodsof DT, DTDS,technique. In this study, the unsupervised classificationand MLL. 1 he unsupervisea IUDATA technique couldmethod of interaclive self-organizing data analysis tech-make signifcant differences and has low accuracy.72Du Jin-kang, FENG Xve-hi, WANG Zhou-long et al.Table 2 The accuracy assessment of four methodsMethodsWATERWATER-WATER+K(%)C(%)DT122048810000DTDS120188318605930259)1MLC .1194441260477814191ISODATA1209325111632716338774 DISCUSSION AND CONCLUSIONREFERENCESIt is difficult to exlract water bodies effectively fromBARTON I J，Bathols J M，1989. Mormitoring foods withAVHRR[J]. Remote Sensing of Environment, 30 (1): 89-SPOT image by applying single technique such as thethresholding in mountainous area due to the efetsof DU Yumyan , ZHOU Chenghu , 198. Atomatcaly Extraet.shadows. The framework of decision tree classifcationing Remote Sensing Information for Waterbodies[J]. Re-could be laken as an effective tool for deniving walermote Sensing of Entironment, 2(4): 264 - 268. (in Chi-bodies because of it's high accuracy, but the designingof clssifer creates dificulties, which requires detailed uU Jian-bo. DAI Chang-da, 1996. The Aplication of TM Im.spectral analysis of the image.age in Reservoir Situation Monitoring[J]. Remote Sensing ofEntironment, 11(1): 53 - 58. (in Chinese)Such auxiliary information as DEM and slope couldLU Jiaju. LI Shi-hong,. 1992. Improvement of the Techniquesbe used to discard shadows from extracted thematic im-for Distinguishing Water Bodies from TM Data[J] . Remoteage of water bodies using thresholding lechnique, butSensing of Enuironment, 7(1):17- 23. (in Chinese)the designed cassifer based on the information of DEM SHENG Yong-wei, XIAO Qian-guang. 1994. Watrbody ldeni-and slope also requires sampling and analysis to find thefication in Cloud-Contamiuted NOAA/ AVHRR Image[J].differences of DEM and slope between pixels of waterRemote Sensing of Entironment, 9(4): 247 -255. (inand other objects.Chinese)The supervised MLC method is an altemative ap-SHIH s F.1985. Comparison of ELAS cassifications and densityslicing Landsat data for waler surface area assessment[A ].proach for it's simple operation procedure and relativelyIn: JOHNSON A I (ed)。 Hydrologic Applications of Spacehigher accuracy. It is easy to perform sample trainingTechnology (Publication No. 160) [C]. Wallingford: IAHSin stretched image of water bodies obtained byPress 91 -97.threshoding approach.XIA0 Qian guang, et ul., 1987. NOAA Imagery Application forThe unsupervised classification method could pro-Monitoring Songhua River Flood[J]. Remote Sensing Infor-duced the results with low accuracy ，which could nolYANG Cun-jian, Xu Mei, 1998. Investigation on Extracting thenation, (4): 26 - 27. (in Chinese)be used as the final products.Waterbody from Landsat TMIJ]. Geographical Reseurch, 17The thresholding technique is not recommended to(supplement): 86 - 89. (in Chinese)be used to obtain the final results in mountainousZHOU Cheng-hu, DU Yun-yan, LUO Jian-cheng, 1996. A De-area, bul it could be used with other methods，there-scription Model Based on Knowledge for Atomatically Rec-fore it is a signifieant method to be used in water ex-ognizing Water from NOAA/AVHRR[J]. Joural of Naturaltraction.Disasters, 5(3): 100- 108. (in Chinese)中国煤化工MYHCNMHG